Better Statistical Machine Translation through Linguistic Treatment of Phrasal Verbs

Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 196–201, October 25-29, 2014, Doha, Qatar. c⃝2014 Association for Computational Linguistics Better Statistical Machine Translation through Linguistic Treatment of Phrasal Verbs Kostadin Cholakov and Valia Kordoni Humboldt-Universit¨at zu Berlin, Germany {kostadin.cholakov,kordonieva}@anglistik.hu-berlin.de Abstract This article describes a linguistically in- formed method for integrating phrasal verbs into statistical machine translation (SMT) systems. In a case study involving English to Bulgarian SMT, we show that our method does not only improve trans- lation quality but also outperforms simi- lar methods previously applied to the same task. We attribute this to the fact that, in contrast to previous work on the subject, we employ detailed linguistic information. We found out that features which describe phrasal verbs as idiomatic or composi- tional contribute most to the better trans- lation quality achieved by our method. 1 Introduction Phrasal verbs are a type of multiword expressions (MWEs) and as such, their meaning is not deriv- able, or is only partially derivable, from the se- mantics of their lexemes. This, together with the high frequency of MWEs in every day communi- cation (see Jackendoff (1997)), calls for a special treatment of such expressions in natural language processing (NLP) applications. Here, we con- centrate on statistical machine translation (SMT) where the word-to-word translation of MWEs of- ten results in wrong translations (Piao et al., 2005). Previous work has shown that the application of dedicated methods to identify MWEs and then integrate them in some way into the SMT pro- cess often improves translation quality. Gener- ally, automatically extracted lexicons of MWEs are employed in the identification step. Further, various integration strategies have been proposed. The so called static strategy suggests training the SMT system on corpora in which each MWE is treated as a single unit, e.g. call off. This im- proves SMT indirectly by improving the align- ment between source and target sentences in the training data. Various versions of this strategy are applied in Lambert and Banchs (2005), Carpuat and Diab (2010), and Simova and Kordoni (2013). In all cases there is some improvement in transla- tion quality, caused mainly by the better treatment of separable PVs, such as in turn the light on. Another strategy, which is referred to as dy- namic, is to modify directly the SMT system. Ren et al. (2009), for example, treat bilingual MWEs pairs as parallel sentences which are then added to training data and subsequently aligned with GIZA++ (Och and Ney, 2003). Other approaches perform feature mining and modify directly the automatically extracted translation table. Ren et al. (2009) and Simova and Kordoni (2013) employ Moses1 to build and train phrase-based SMT sys- tems and then, in addition to the standard phrasal translational probabilities, they add a binary fea- ture which indicates whether an MWE is present in a given source phrase or not. Carpuat and Diab (2010) employ the same approach but the additional feature indicates the number of MWEs in each phrase. All studies …

Content truncated to 500 words